CoolFace
Modelpublic

Brandi-Kinard/plainspeak-smollm2-1.7b

sourceHugging Facemitupdated 5mo agoView on Hugging Face
0likes23downloads
Model Card

PlainSpeak — Dense-to-Plain-English Translator

I taught a tiny AI to speak human.

Give it Shakespeare. Give it a legal contract. Give it anything written to impress instead of communicate. It gives you back what it actually means.

Try it

[→ Live demo on Hugging Face Spaces](https://huggingface.co/spaces/Brandi-Kinard/plainspeak)

bash
pip install mlx-lm
mlx_lm.generate \
  --model Brandi-Kinard/plainspeak-smollm2-1.7b \
  --prompt "### Original:
YOUR TEXT HERE

### Plain English:" \
  --max-tokens 200

Examples

Shakespeare → Plain English

Original: "Wherefore art thou Romeo? Deny thy father and refuse thy name." PlainSpeak: "Why are you Romeo? Don't deny your father and refuse your name."

Adam Smith → Plain English

Original: "The invisible hand of the market, whereby individuals pursuing their own self-interest are led, as if by an invisible hand, to promote ends which were no part of their original intention." PlainSpeak: "When people try to make money for themselves, they often end up helping society without meaning to."

KJV Bible → Plain English

Original: "The LORD is my shepherd; I shall not want." PlainSpeak: "The LORD leads me. I don't need anything else."

Model Details

PropertyValue
Base modelSmolLM2-1.7B-Instruct
Fine-tuning methodLoRA (8 layers)
Training iterations500
Training examples1,200
Validation examples150
Data sourceProject Gutenberg + AI-generated synthetic pairs
HardwareApple M1, 16GB unified memory
Peak training memory10.09 GB
Final val loss1.771
Inference memory~3.6 GB
Build time1 evening

What it's good at

  • —19th century prose (Dickens, James, Eliot, Hardy)
  • —Shakespeare and Elizabethan English
  • —King James Bible passages
  • —Economic and political theory (Smith, Burke, Locke)
  • —Academic abstracts
  • —Legal boilerplate

Known limitations

  • —Trained on 200-word chunks — short fragments may produce inconsistent results
  • —Occasional errors on numerical content (dates, quantities)
  • —Not optimized for highly technical scientific notation
  • —May struggle with extremely abstract or experimental writing (e.g. stream-of-consciousness)

How it was built

1. Stream 1,500 prose passages from Project Gutenberg
2. Generate plain English versions using a frontier model as teacher
3. Format as (original → plain) training pairs
4. Fine-tune SmolLM2-1.7B with LoRA on Apple MLX
5. Fuse adapter into final weights

The key insight: a small model trained on 1,500 high-quality examples outperforms a large model trained on millions of noisy ones — at this specific task.

Use in Python

python
from mlx_lm import load, generate

model, tokenizer = load("Brandi-Kinard/plainspeak-smollm2-1.7b")

prompt = """### Original:
It is a truth universally acknowledged, that a single man in possession of a good fortune, must be in want of a wife.

### Plain English:"""

response = generate(model, tokenizer, prompt=prompt, max_tokens=200)
print(response)

Links